[Bug tree-optimization/70729] Loop marked with omp simd pragma is not vectorized

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70729

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2016-04-19
     Ever confirmed|0                           |1

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
So the argument is that C1[S_n+1] doesn't alias in[] because of the omp simd
pragma and its guarantees?  Otherwise I don't see how that can be guaranteed.

loop->safelen is 2147483647

it seems that INT_MAX is a special value (of course loops could iterate
more than INT_MAX times, so that special value is badly choosen - better
use some negative value like -1 as special value).

And invariant motion could use that property but does not.

Index: gcc/tree-ssa-loop-im.c
===================================================================
--- gcc/tree-ssa-loop-im.c      (revision 235201)
+++ gcc/tree-ssa-loop-im.c      (working copy)
@@ -2198,6 +2198,9 @@ ref_indep_loop_p (struct loop *loop, im_
 {
   gcc_checking_assert (MEM_ANALYZABLE (ref));

+  if (loop->safelen == INT_MAX)
+    return true;
+
   return ref_indep_loop_p_2 (loop, ref, false);
 }


fixes this.  But better abstract loop->safelen == INT_MAX into a predicate
in cfgloop.h.

But even with that the loop is not vectorized because of

t.C:66:25: note: def_stmt: prephitmp_56 = PHI <_5(4), _20(6)>
t.C:66:25: note: type of def: unknown
t.C:66:25: note: Unsupported pattern.
t.C:66:25: note: not vectorized: unsupported use in stmt.
t.C:66:25: note: unexpected pattern.

because even if LIM manages to hoist this PRE has messed up this before
and we end up with

  <bb 2>:
  _5 = this_4(D)->S_n;
  if (_5 > 0)
    goto <bb 4>;
  else
    goto <bb 3>;

  <bb 3>:
  return;

  <bb 4>:
  pretmp_54 = this_4(D)->C2;
  pretmp_57 = this_4(D)->C1;
  pretmp_60 = MEM[(int * *)this_4(D) + 56B];
  _20 = this_4(D)->S_n;

  <bb 5>:
  # i_33 = PHI <0(4), i_28(6)>
  # prephitmp_56 = PHI <_5(4), _20(6)>

notice the redundant this->S_n load.